AI usage — how much it costs and who uses it

The Usage report answers three questions: how many requests, how many tokens, and how much did it cost.

Step by step

Open the report and set the date range

1. In the sidebar, under Dashboard, expand Reports → AI and select Usage. This requires permission for the AI Usage report (the Report group in field permissions).
2. The screen is titled AI Usage Statistics, with the subtitle Monitor AI integration usage and costs. The information icon beside the title lists exactly what is counted.
3. Set Date range to Last 7 days, Last 30 days (default), Last 90 days, This year, or Custom. With Custom, the From and To fields appear.
4. Narrow the results with the Model list (defaults to All models) and the Context list (All contexts).
5. Refresh recalculates the data. The report is not cached, so a broad date range can take longer to load. Beside the button you can see the exact range and number of days included. All charts and tables use the same range, model, and context.

Find what the money is spent on and by whom

1. Start with the six tiles at the top: Requests, Success rate, Tokens, Total cost, Avg. time, and Errors. Amounts are in dollars. Under the cost, the report indicates whether the provider supplied the actual charge or whether the amount is an estimate from older data. Success rate is calculated for completed requests; Errors also include timeouts.
2. The Requests per day, Tokens per day, and Daily cost charts show whether costs are rising steadily or in spikes.
3. Usage by model answers whether an expensive model is doing work a cheaper one could handle. Usage by context shows which system feature is using it.
4. Cost per user is a table with User, Requests, Successes, Errors, Tokens, Cost, and Avg. time columns. One user with a disproportionate share usually indicates a series of questions. If there is no data, the report says No user usage data.
5. Recent requests has Model, Context, Tokens, Cost, Time, Status, and Created columns. Status and Context are shown as raw database values, in English.
6. The eye icon (View details) opens the AI Request Details dialog: Model, Context, Response time, Cost, and below those Input tokens, Output tokens, and All tokens. If content was saved, it also shows System prompt, User prompt, and AI response. Otherwise, it says Full request/response details are unavailable for this request. Close it with Close.
7. The report is read-only and does not require 2FA confirmation.

When the screen is empty

1. No AI usage data, with the text AI usage statistics will appear here once you start using AI features and a Configure AI integration button, usually just means there were no requests in the selected period.
2. If data could not be loaded, the report shows a separate red error message instead of pretending there are no results. Check the connection and whether the role has Read access to the Reports module and field permission for AI Usage.

What is on the screen

• Summary metrics — number of requests, tokens, and cost.
• Daily charts — requests, tokens, and cost over time.
• Model breakdown — usage by model.
• User table — how much each person used.
• Recent requests — with a view of each individual request.

Ranges: 7, 30, and 90 days, the current year, or a custom range. The default is 30 days.

What it is useful for

• Cost control. The model breakdown shows whether an expensive model is being used for tasks a cheaper one could handle.
• Abuse detection. One person with a disproportionate share usually means repeated questions or automation running under that user’s account.
• Justifying limits. The rate limit is shared across the company, so one intensive user slows everyone else down (Configuring the assistant — integration, model, prompts, and costs).

Data is calculated live

This report is not cached — the numbers are current when you open it. With a large date range, it may load more slowly than logistics reports.

Where the costs come from

For every assistant response, the system records tokens and the cost returned by the provider. If an older or only partially compatible integration does not provide a cost, the system uses a local estimate and marks it in the Total cost tile (Configuring the assistant — integration, model, prompts, and costs).

Note: an interrupted question can still cost money. If the model calculated an answer before the interruption, the charge was incurred even though no one saw the answer (Assistant limitations and common misunderstandings).